fix(sandbox): keep programmatic tool calling items in memory rollouts

[hsusul]

Summary

Sandbox memory rollouts silently drop Programmatic Tool Calling items.

_should_include_memory_item() in src/agents/sandbox/memory/rollouts.py keeps an item only when its type is listed in _INCLUDED_MEMORY_ITEM_TYPES (an explicit _EXCLUDED_MEMORY_ITEM_TYPES set names the types that are deliberately dropped; everything else falls through to dropped). That set lists both halves of every call/output pair in _TOOL_CALL_TO_OUTPUT_TYPE (run_internal/items.py) — function_call, custom_tool_call, shell_call, apply_patch_call, computer_call, local_shell_call, tool_search_call and their outputs — plus web_search_call, mcp_call, mcp_approval_request, mcp_approval_response.

The one pair missing is program / program_output.

The set was written in #2889 (2026-04-15) and has not been edited since. Programmatic Tool Calling added the program and program_output item types later, in #3833 (2026-07-17), and this call site was not updated — so a run that uses ProgrammaticToolCallingTool writes a rollout with the program code and the program result removed.

The rollout is the raw evidence that phase-one/phase-two memory generation reads, so the loss is not cosmetic. Worse, the filter is inconsistent with itself: the function_call items a program issues are retained, and each one carries caller: {"type": "program", "caller_id": ...} pointing at a program item that was just removed. The stored rollout therefore contains dangling caller references.

Before, for a single programmatic turn (program → function_call → function_call_output → program_output):

generated_items: ["function_call", "function_call_output"]
  function_call.caller = {"type": "program", "caller_id": "call_prog_1"}   # target not in the rollout

After:

generated_items: ["program", "function_call", "function_call_output", "program_output"]

Fix: add "program" and "program_output" to the existing frozenset. Two strings — no new branch, no new helper, no public API or stored-format change.

Deliberately not widened. file_search_call, code_interpreter_call, mcp_list_tools and hosted_tool_call are also absent from the set, but they are not part of the contract this change asserts: the claim here is only that every call/output pair the SDK knows how to pair should survive into the rollout, and program is the sole pair that does not. image_generation_call stays excluded via _EXCLUDED_MEMORY_ITEM_TYPES. There is a boundary test covering exactly this.

Test plan

Three tests added to tests/sandbox/test_memory.py:

• test_build_rollout_payload_keeps_programmatic_tool_calling_items — a full programmatic turn as RunItems (Program, a function_call carrying CallerProgram, its output, ProgramOutput). Asserts all four survive in order, that the program's code and the program output's result are preserved, and that the retained function_call.caller resolves to a program item that is present.
• test_build_rollout_payload_keeps_program_items_from_input — the same types arriving on the input side, which goes through the same filter.
• test_build_rollout_payload_still_drops_hosted_items_outside_the_included_set — boundary: file_search_call and image_generation_call are still dropped while program is kept, so the narrow scope is pinned.

All three fail on upstream/main @ f3b6c617 and pass with the change; verified by reverting only the source hunk and re-running (3 failed, 3 passed → 6 passed). The three pre-existing build_rollout_payload tests pass unchanged in both states.

Commands run:

• uv run pytest tests/sandbox/test_memory.py -q → 73 passed
• uv run pytest tests/sandbox tests/test_programmatic_tool_calling.py -q → 1130 passed, 2 skipped
• make format → 862 files left unchanged
• make lint → All checks passed!
• make typecheck → mypy and pyright clean (exit 0)
• make tests → 6744 passed, 29 skipped (parallel) + 77 passed, 5 skipped (serial)
• git diff --check → clean

No API key, no network, no live service.

Not run: make coverage, make build-docs, the Python 3.10 matrix, and the integration-test profiles.

One unrelated flake seen while validating: tests/extensions/memory/test_advanced_sqlite_session.py::test_branch_allocation_is_serialized_across_processes fails intermittently (~1 in 5) under CPU contention at assert first_ready.wait(timeout=10), a fixed wall-clock wait on a multiprocessing.Event. It reproduces with this change reverted and is not in this change's import graph; make tests is green on this branch when the machine is idle.

Issue number

None.

Checks

• I've added new tests, if relevant
• I've run .agents/skills/code-change-verification/scripts/run.sh
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[seratch]

Can you resolve the typecheck error?